Validation of 360-Degree Camera Data
Anonymized for review
Validating ecological data extracted from 360-degree images using the pannotator package with field survey data
This file contains a preliminary assessment of the ability of data extracted from 360-degree panospheric images using the pannotator software to reliably capture important ecological variables that have been reported in the accompanying paper. The analyses and results reported in this document are based on data collected over a series of validation plots located across Uluru-Kata Tjuta National Park, Australia in which field survey data and 360-degree images were collected and in the main study transect reported in the paper.
Here we focus on three sets of analyses that investigate the reliability and repeatability of key ecological data presented in the paper:
A comparison of plot-level plant species lists constructed from both 360-degree camera images and accompanying field survey data (based on validation plots)
A comparison of plant cover estimates, again using both 360-degree camera images and field survey data (based on validation plots)
Consistency in the scoring of tree crown health by different users of the pannotator package (based on the main transect)
We note that the analyses presented here relate only to the data collected in our study, and that research focusing on different species or systems will likely require different methods for collecting 360-degree images and for calibration. However, based on our results we draw general conclusions relevant to other study systems and provide practical advice for collecting images and calibrating image-based data in the accompanying paper.
Data sources
The species and cover analyses conducted below are based on data collected from a series of validation plots (20 m diameter) located on the edges or corners of square transects that measure 1 km on each side. The plots were established across Uluru-Kata Tjuta National Park to capture variation in landforms and vegetation across the Park. Crown data for desert oak (Allocasuarina decaisneana) were extracted from 360-degree images collected along the main 4 km long study transect centred on S 25.35 degrees, W 130.72 degrees (see accompanying paper).
Install the necessary packages for all code below
Attaching package: 'dplyr'
The following objects are masked from 'package:stats':
filter, lag
The following objects are masked from 'package:base':
intersect, setdiff, setequal, union
Attaching package: 'data.table'
The following objects are masked from 'package:dplyr':
between, first, last
Linking to GEOS 3.11.2, GDAL 3.6.2, PROJ 9.2.0; sf_use_s2() is TRUE
── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
✔ forcats 1.0.0 ✔ purrr 1.0.2
✔ lubridate 1.9.3 ✔ tibble 3.2.1
── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
✖ data.table::between() masks dplyr::between()
✖ dplyr::filter() masks stats::filter()
✖ data.table::first() masks dplyr::first()
✖ lubridate::hour() masks data.table::hour()
✖ lubridate::isoweek() masks data.table::isoweek()
✖ dplyr::lag() masks stats::lag()
✖ data.table::last() masks dplyr::last()
✖ lubridate::mday() masks data.table::mday()
✖ lubridate::minute() masks data.table::minute()
✖ lubridate::month() masks data.table::month()
✖ lubridate::quarter() masks data.table::quarter()
✖ lubridate::second() masks data.table::second()
✖ purrr::transpose() masks data.table::transpose()
✖ lubridate::wday() masks data.table::wday()
✖ lubridate::week() masks data.table::week()
✖ lubridate::yday() masks data.table::yday()
✖ lubridate::year() masks data.table::year()
ℹ Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errors1. Plant species data
All woody plant species (subshrubs < 50 cm tall, shrubs >= 50 cm tall and trees) along with species of Triodia (spinifex grasses) were identified inside 79 validation plots (20 m diameter) during field surveys conducted in August 2022. Here we read in the .csv file and create the R dataframe ‘species_data’.
The dataframe ‘species_data’ contains the following variables:
Column \[1\] (ID) = record identifier; \[2\] (Plot) = plot number, \[3\] (Location) = location of plot along transect (NE, NW, SE, SW corner or extra plot); \[4\] (Latitude) = latitude in decimal degrees; \[5\] (Longitude) = longitude in decimal degrees; \[6\] (Closest_image) = identity of closest 360-degree image (for pannotator annotation); \[7\] (Camera_d_to_plot) = approximate distance (m) from camera image centre to survey plot centre; \[8\] (No_Field_species) = number of species identified in field survey; \[9\] (No_Camera_species) = number of species identified in camera survey; \[10\] (In_field_not_camera) = number of species unique to field survey;
\[11\] (Notes) = notes on errors in data collection, where ‘NoDiscrepancy’ = no differences between field survey and camera survey species lists, ‘Inside5m’ = the species not identified (missed in the survey) was within 5 m of the plot centre and camera, ‘Outside5m’ = the species missed was more than 5m from the plot centre and camera, and ‘FieldError’ = the species was incorrectly recorded as absent in the field survey;
\[12\]-\[14\] (Field_extra_Species_1, Field_extra_Species_2, Field_extra_Species_3) = names of extra species recorded in the field survey; \[15\] (In_camera_not_field) = number of species unique to camera survey; \[16\] (C_E_Species_1) = names of extra species recorded in the camera survey; \[17\] (Shared_species) = total number of shared species in field and camera survey lists; \[18\] (Unique_species) = total number of unique species in field and camera survey lists; \[19\] (Agreement) = Agreement (%), measured as 100 X ((shared species) / (shared + unique species));
\[20\]-\[51\] (AcacAneu, AlloDeca, Triodiasp, AlutMais, GrevErio, AcacLigu, EucaGamo, DicrGile, AcacPrui, EremSpec, DodoVisc, AcacMell, CodoCoti, GyroTepp, AcacMiny, EucaOxy, NewcSpod, AcacMait, ExocSpar, HakeSpec, GrevSten, SanLanc, SennArte, GyroRamo GrevJunc, AcacAmmo, EremLatr, EremGlab, EremGibs, RhagErem, EremLong, Acacia_sp) species present in plots, with Both = species in both field and camera survey lists; CameraOnly = species found only in camera survey list; FieldOnly = species found only in field survey list; Absent = species not recorded in either field or camera survey data.
The species are: AcacAneu = Acacia aneura, AlloDeca = Allocasuarina decaisneana, Triodiasp = Triodia species (T. pungens, T. schinzii and T. basedowii), AlutMais = Aluta maisonneuvei, GrevErio = Grevillea eriostachya, AcacLigu = Acacia ligulata, EucaGamo = Eucalyptus gammophylla, DicrGile = Dicrastylis gilesii, AcacPrui = Acacia pruinocarpa, EremSpec = Eremophila sp., DodoVisc = Dodonaea viscosa, AcacMell = Acacia melliodora, CodoCoti = Codonocarpus cotinifolius, GyroTepp = Gyrostemon tepperi, AcacMiny = Acacia minyura, EucaOxy = Eucalyptus oxymitra, NewcSpod = Newcastelia spodiotricha, AcacMait = Acacia maitlandii, ExocSpar = Exocarpos sparteus, HakeSpec = Hakea species , GrevSten = Grevillea stenobotrya, SanLanc = Santalum lanceolatum, SennArte = Senna artemisiodes, GyroRamo = Gyrostemon ramulosum, GrevJunc = Grevillea juncifolia, AcacAmmo = Acacia ammobia , EremLatr = Eremophila latrobei, EremGlab = Eremophila glabra, EremGibs = Eremophila gibsonii , RhagErem = Rhagodia eremaea , EremLong = Eremophila longifolia, Acacia_sp = Acacia species.
species_data <- read_csv("./Data_files/validation_species.csv", show_col_types = FALSE)
# confirm that there are 79 plots of species data
cat("The number of rows in the dataframe is: ", nrow(species_data))
# initialize table counter
table_counter <- 0
# function to generate table numbers and captions
table_caption <- function(caption) {
table_counter <<- table_counter + 1
paste("Table", table_counter, ":", caption)
}
# show the first 10 rows of the dataframe
#kable(head(species_data, 10), format = "html", caption="Species calibration data frame") The number of rows in the dataframe is: 79We begin by comparing the mean species number across plots using species lists compiled from field survey data and camera images scored in the pannotator package.
# find the mean number of species in the field surveys
Field_mean <- summary(species_data$No_Field_species)
cat("The summary for the number of species based on field survey data is:\n")
print(Field_mean)
cat("\n")
# find the mean number of species in the camera surveys
Camera_mean <- summary(species_data$No_Camera_species)
cat("The summary for the number of species based on camera image data is:\n")
print(Camera_mean)
cat("\n")The summary for the number of species based on field survey data is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
1.000 2.000 3.000 3.266 5.000 7.000
The summary for the number of species based on camera image data is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
1.000 2.000 3.000 2.861 4.000 6.000 The data show that on average we recorded 3.3 species per plot based on field surveys and 2.9 species based on camera images. Thus, there was a weak tendency to record fewer species using camera images. Now we determine the relationship between plot-level species counts in the field survey versus camera survey counts.
# abline adds a linear model to the plot.
plot_sp_richness <- plot(species_data$No_Field_species, species_data$No_Camera_species,
main = "Plot-level species richness",
xlab = "No. of species (Field Survey)",
ylab = "No. of species (Camera Survey)",
pch = 16, # Use filled circles as data points
col = "black", # Set point color
ylim = c(0, 8), # Set y-axis limits
xlim = c(0, 8)) # Set x-axis limits
abline(lm(No_Camera_species ~ No_Field_species, data = species_data), col = "red")
# create linear model for species data
model <- lm(No_Camera_species ~ No_Field_species, data = species_data)
cat("Model results:\n")
print(model)
# summarise linear model
model_summary <- summary(model)
cat("Model summary:\n")
print(model_summary)Model results:
Call:
lm(formula = No_Camera_species ~ No_Field_species, data = species_data)
Coefficients:
(Intercept) No_Field_species
0.4394 0.7414
Model summary:
Call:
lm(formula = No_Camera_species ~ No_Field_species, data = species_data)
Residuals:
Min 1Q Median 3Q Max
-1.14654 -0.18080 0.07777 0.33633 1.33633
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.43936 0.14054 3.126 0.0025 **
No_Field_species 0.74144 0.03885 19.087 <2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.5374 on 77 degrees of freedom
Multiple R-squared: 0.8255, Adjusted R-squared: 0.8233
F-statistic: 364.3 on 1 and 77 DF, p-value: < 2.2e-16There is a strong positive relationship between species counts in plots based on field survey and camera survey data which explains 82.3 % of the variation in the data (p < 6.3825904^{-31}). The slope (0.74) again indicates a weak tendency for species lists compiled using camera imagery to be smaller than those based on field survey data.
Now we assess the percentage agreement in the species lists based on field survey and camera survey data. The percentage agreement (A) is measured as A = 100 X ((shared species) / (shared + unique species)).
Mean_agreement = summary(species_data$Agreement)
cat("The summary for % agreement between field & camera survey species lists is:\n")
print(Mean_agreement)
# construct a histogram of the % agreement between mean plot-level species counts based on field and camera surveys
hist_data <- hist(species_data$Agreement, breaks = seq(0, 100, by = 5),
main = "Histogram of Field-Camera survey agreement",
xlab = "Percent agreement",
ylab = "Frequency")
The summary for % agreement between field & camera survey species lists is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
25.00 66.67 100.00 83.75 100.00 100.00 On average there was 84% agreement between field survey and camera survey species lists. The histogram shows that a majority of plots were in 100% agreement, with a second cluster at 65-90% agreement. A small number of plots had lower agreement.
We now investigate which species were contributing to the lack of agreement in species lists in certain plots. First we construct a list of the species that were missed in the camera surveys of each plot (i.e., were found only in the field survey), with the number of occasions of each. There are three columns of species found only in the field survey, labelled Field_extra_Species_1, Field_extra_Species_2, and Field_extra_Species_3. Next we construct a list of the species that were recorded ONLY in the camera surveys of each plot (i.e., were not recorded in the field survey), with counts of each. There is one column of species found only in the camera survey, labelled C_E_Species_1.
# summarise results for each of the field extra species columns
summary_1 <- species_data %>%
filter(!is.na(Field_extra_Species_1)) %>%
count(Field_extra_Species_1, sort = TRUE) %>%
rename(Species = Field_extra_Species_1) %>%
mutate(Species = as.character(Species))
summary_2 <- species_data %>%
filter(!is.na(Field_extra_Species_2)) %>%
count(Field_extra_Species_2, sort = TRUE) %>%
rename(Species = Field_extra_Species_2) %>%
mutate(Species = as.character(Species))
summary_3 <- species_data %>%
filter(!is.na(Field_extra_Species_3)) %>%
count(Field_extra_Species_3, sort = TRUE) %>%
rename(Species = Field_extra_Species_3) %>%
mutate(Species = as.character(Species))
field_only_species <- bind_rows(summary_1, summary_2, summary_3) %>%
group_by(Species) %>%
summarise(n = sum(n)) %>%
arrange(desc(n))
# print table using gt
field_only_species_gt_table <- gt(field_only_species) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20),
quarto.use_bootstrap = TRUE
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Extra species in field surveys")
)
field_only_species_gt_table| Table 1 : Extra species in field surveys | |
|---|---|
| Species | n |
| Triodia_schinzii | 11 |
| Acacia_ligulata | 4 |
| Aluta_maisonneuvei | 3 |
| Acacia_melliodora | 2 |
| Acacia_minyura | 2 |
| Eremophila_glabra | 2 |
| Eremophila_latrobei | 2 |
| Eremophila_longifolia | 2 |
| Grevillea_juncifolia | 2 |
| Acacia sp. | 1 |
| Acacia_maitlandii | 1 |
| Codonocarpus_cotinifolius | 1 |
| Dodonea_viscosa | 1 |
| Eucalyptus_oxymitra | 1 |
| Grevillea_eriostachya | 1 |
| Grevillea_stenobotrys | 1 |
| Hakea_?_lorea | 1 |
| Rhagodia_eremaea | 1 |
| Santalum_lanceolatum | 1 |
| Senna_artemisiodes | 1 |
# summarise results for each of the species recorded in camera images only
camera_only_species <- species_data %>%
filter(!is.na(C_E_Species_1)) %>%
count(C_E_Species_1, sort = TRUE) %>%
rename(Species = 1) %>%
mutate(Species = as.character(Species))
camera_only_species_gt_table <- gt(camera_only_species) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20),
quarto.use_bootstrap = TRUE
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Extra species in camera surveys")
)
camera_only_species_gt_table| Table 2 : Extra species in camera surveys | |
|---|---|
| Species | n |
| Aluta_maisonneuvei | 4 |
| Acacia_aneura | 3 |
| Triodia_sp. | 3 |
| Acacia_maitlandii | 1 |
| Grevillea_eriostachya | 1 |
| Newcastelia_spodiotricha | 1 |
These data show that the most common species identified only in the field survey of a given plot was Triodia schinzii, a spinifex grass. In many plots Triodia pungens and Triodia schinzii were both present, but it was not generally possible to distinguish these taxa, the difference of which can usually only be determined by panicle or floral structure. Acacia ligulata and Aluta maisonneuvei were also missed in the camera surveys 4 and 3 times respectively, but these were often in burnt plots. Other species, usually shrubs, were occasionally missed.
We now determine some circumstances under which species were not recorded in the two surveys. For this we have the variable species_data$Notes, which has the categories ‘NoDiscrepancy’ = the field and camera survey lists were the same, ‘Outside5m’ = detected in field survey 5 m to 10 m from the plot centre but not in the camera survey, ‘Inside5m’ = detected in field survey within 5 m of the plot centre but not in the camera survey, and ‘FieldError’ where an error was made in the field survey.
# count occurrences in 'Notes' column and sort the results
df <- species_data %>%
count(Notes, sort = TRUE)
# calculate the total number of records
total <- sum(df$n)
# calculate the total number of discrepancies in plots species lists
# need to filter by specific Notes values corresponding to discrepancies
disc <- df %>%
filter(Notes %in% c("Inside5m", "Outside5m", "FieldError")) %>%
summarise(total_discrepancies = sum(n)) %>%
pull(total_discrepancies)
# calculate the percentage of discrepancies
pcnt_disc <- 100 * (disc / total)
# print the percentage of discrepancies
cat(paste("Percentage of discrepancies:", round(pcnt_disc, 1), "%"))
cat("\n")
df_gt_table <- gt(df) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Discrepancies between field & camera surveys")
)
df_gt_tablePercentage of discrepancies: 45.6 %| Table 3 : Discrepancies between field & camera surveys | |
|---|---|
| Notes | n |
| NoDiscrepancy | 43 |
| Outside5m | 22 |
| Inside5m | 11 |
| FieldError | 3 |
These data show that there was discrepancy in species lists generate from field and camera surveys in 46% of plots (in 36 of 79 plots), and that in 22 plots (28%) at least one species located more than 5 m from the centre of the plot was missed in the camera survey. In 11 plots (14%) a species was missed in the camera survey when it occurred less than 5 m from the centre of the plot.
We now determine which species were missed < 5 m and > 5 m from the plot centre. Those missed < 5 m from the plot centre are the most serious because they fall within the area used to determine species counts in subsequent camera-based surveys (see accompanying paper).
# Filtering species_data for notes "Inside5m" and removing rows with NA's in Field_extra_Species_1 and Field_extra_Species_2
species_data_L5 <- species_data %>%
filter(Notes == "Inside5m") %>%
filter(complete.cases(Field_extra_Species_1))
# Get the species counts for Field_extra_Species_1
summary_1_L5 <- species_data_L5 %>%
count(Field_extra_Species_1, sort = TRUE) %>%
rename(Species = Field_extra_Species_1) %>%
mutate(Species = as.character(Species))
# Get the species counts for Field_extra_Species_2
summary_2_L5 <- species_data_L5 %>%
filter(!is.na(Field_extra_Species_2)) %>%
count(Field_extra_Species_2, sort = TRUE) %>%
rename(Species = Field_extra_Species_2) %>%
mutate(Species = as.character(Species))
# Combine the counts from both columns
Less5_summary <- bind_rows(summary_1_L5, summary_2_L5) %>%
group_by(Species) %>%
summarise(n = sum(n)) %>%
arrange(desc(n))
Less5_summary_gt_table <- gt(Less5_summary) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(title = table_caption("Species missed inside 5 metres"))
Less5_summary_gt_table
# get the total count of missed Triodia schinzii < 5 m from the plot center
Triodia_index <- which(Less5_summary$Species == "Triodia_schinzii")
Triodia_count <- Less5_summary$n[Triodia_index]
# using `summarize` to calculate the total count from Less5_summary
Total_count <- Less5_summary %>%
summarize(total = sum(n)) %>%
pull(total)
# calculate the percentage of Triodia schinzii
Triodia_pcnt <- round(100 * (Triodia_count/Total_count),1)
cat(paste("Missed Triodia schinzii (< 5 m) as % of total: ", Triodia_pcnt, "%"))
cat("\n")| Table 4 : Species missed inside 5 metres | |
|---|---|
| Species | n |
| Triodia_schinzii | 7 |
| Acacia_ligulata | 1 |
| Acacia_melliodora | 1 |
| Eremophila_glabra | 1 |
| Eucalyptus_oxymitra | 1 |
| Rhagodia_eremaea | 1 |
Missed Triodia schinzii (< 5 m) as % of total: 58.3 %Now do the same for species greater than 5 m
# Filtering species_data for notes "Outside5m" and removing rows with NA's in Field_extra_Species_1 and Field_extra_Species_2
species_data_Gr5 <- species_data %>%
filter(Notes == "Outside5m") %>%
filter(complete.cases(Field_extra_Species_1))
# Get the species counts for Field_extra_Species_1
summary_1_Gr5 <- species_data_Gr5 %>%
count(Field_extra_Species_1, sort = TRUE) %>%
rename(Species = Field_extra_Species_1) %>%
mutate(Species = as.character(Species))
# Get the species counts for Field_extra_Species_2
summary_2_Gr5 <- species_data_Gr5 %>%
filter(!is.na(Field_extra_Species_2)) %>%
count(Field_extra_Species_2, sort = TRUE) %>%
rename(Species = Field_extra_Species_2) %>%
mutate(Species = as.character(Species))
# Combine the counts from both columns
Gr5_summary <- bind_rows(summary_1_Gr5, summary_2_Gr5) %>%
group_by(Species) %>%
summarise(n = sum(n)) %>%
arrange(desc(n))
Gr5_summary_gt_table <- gt(Gr5_summary) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Species missed outside 5 metres")
)
Gr5_summary_gt_table| Table 5 : Species missed outside 5 metres | |
|---|---|
| Species | n |
| Triodia_schinzii | 4 |
| Aluta_maisonneuvei | 3 |
| Acacia_ligulata | 2 |
| Acacia_minyura | 2 |
| Eremophila_latrobei | 2 |
| Eremophila_longifolia | 2 |
| Acacia sp. | 1 |
| Acacia_maitlandii | 1 |
| Acacia_melliodora | 1 |
| Codonocarpus_cotinifolius | 1 |
| Dodonea_viscosa | 1 |
| Eremophila_glabra | 1 |
| Grevillea_eriostachya | 1 |
| Grevillea_juncifolia | 1 |
| Grevillea_stenobotrys | 1 |
| Hakea_?_lorea | 1 |
| Santalum_lanceolatum | 1 |
| Senna_artemisiodes | 1 |
For species missed within 5 m of the plot centre, 7/12 cases 58.3% were of Triodia schinzii (which could not be distinguished easily from Triodia pungens, especially when sterile). Other species were only missed once. A much greater number of species were missed 5 m to 10 m from the plot centre, most being small shrubs or seedlings, burnt plants that retained no leaf material, or species that are morphologically similar to others in the study area (e.g., Triodia schinzii , Acacia minyura, various Eremophila spp.). Collectively these data suggest that given the resolution of GoPro Max cameras shrubby and smaller species should be recorded within 5 m of the plot centre unless they are morphologically or chromatically distinct (e.g., Allocasuarina descaisneana, Corymbia opaca).
The main species recorded in the camera survey but not in the field (within individual plots) included Aluta maisonneuvei, Acacia aneura, and Triodia sp. (spinifex grasses). Acacia aneura is morphologically similar to Acacia minyura, and burnt Aluta maisonneuvei could be mistaken for other shrub species. However, some field survey errors also occurred, particularly where Tridodia spp. were overlooked.
We now complete this analysis by determining the field detection rate of plant species recorded across all study plots, later focusing only on widespread species recorded in at least ten plots. For the following analysis we now select only the record ID (plot) and record of each definitively identified species (including the 3 Triodia spp. species lumped) in a plot, where ‘Both’ = species recorded in both field and camera survey lists; ‘CameraOnly’ = species recorded only in camera survey list; ‘FieldOnly’ = species recorded only in field survey list; ‘Absent’ = species not recorded in either field or camera survey data.
# take a subset of the species_data
dataB <- species_data[, c("ID", "AcacAneu", "AlloDeca", "Triodiasp", "AlutMais", "GrevErio", "AcacLigu", "EucaGamo", "DicrGile", "AcacPrui", "CodoCoti", "DodoVisc", "GyroTepp", "AcacMell", "AcacMiny", "EucaOxy", "NewcSpod","AcacMait","ExocSpar", "GrevSten", "SanLanc", "GrevJunc", "SennArte", "GyroRamo", "AcacAmmo", "EremLatr", "EremGlab", "RhagErem", "EremLong")]
# some data wrangling to produce a pivot table for each species based on numbers of 'Both', 'CameraOnly', 'FieldOnly', and 'Absent' identifications in each plot
dataC <- dataB %>%
pivot_longer(cols = -ID) %>%
count(name, value) %>%
pivot_wider(names_from = name, values_from = n, values_fill = list(n = 0))
# transpose the table
t_dataC <- t(dataC)
#get row and colnames in order
colnames(t_dataC) <- rownames(dataC)
rownames(t_dataC) <- colnames(dataC)
#View(t_dataC)
# assign the names
colnames(t_dataC) <- t_dataC[1, ]
t_dataC2 <- t_dataC[-1,]
# convert to data frame and add row names as a column
species_counts <- as.data.frame(t_dataC2) %>%
rownames_to_column(var = "Species")
#View(species_counts)
species_counts_gt_table <- gt(species_counts) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Species counts based on numbers of 'Both', 'CameraOnly', 'FieldOnly', and 'Absent' identifications in each plot")
)
species_counts_gt_table| Table 6 : Species counts based on numbers of ‘Both’, ‘CameraOnly’, ‘FieldOnly’, and ‘Absent’ identifications in each plot | ||||
|---|---|---|---|---|
| Species | Absent | Both | CameraOnly | FieldOnly |
| AcacAmmo | 77 | 2 | 0 | 0 |
| AcacAneu | 50 | 26 | 3 | 0 |
| AcacLigu | 68 | 7 | 0 | 4 |
| AcacMait | 75 | 2 | 1 | 1 |
| AcacMell | 77 | 0 | 0 | 2 |
| AcacMiny | 73 | 5 | 0 | 1 |
| AcacPrui | 78 | 1 | 0 | 0 |
| AlloDeca | 59 | 20 | 0 | 0 |
| AlutMais | 52 | 20 | 4 | 3 |
| CodoCoti | 66 | 12 | 0 | 1 |
| DicrGile | 78 | 1 | 0 | 0 |
| DodoVisc | 76 | 2 | 0 | 1 |
| EremGlab | 77 | 0 | 0 | 2 |
| EremLatr | 75 | 2 | 0 | 2 |
| EremLong | 77 | 2 | 0 | 0 |
| EucaGamo | 71 | 8 | 0 | 0 |
| EucaOxy | 76 | 2 | 0 | 1 |
| ExocSpar | 78 | 1 | 0 | 0 |
| GrevErio | 59 | 18 | 1 | 1 |
| GrevJunc | 77 | 0 | 0 | 2 |
| GrevSten | 76 | 2 | 0 | 1 |
| GyroRamo | 78 | 1 | 0 | 0 |
| GyroTepp | 77 | 2 | 0 | 0 |
| NewcSpod | 75 | 3 | 1 | 0 |
| RhagErem | 78 | 0 | 0 | 1 |
| SanLanc | 78 | 0 | 0 | 1 |
| SennArte | 77 | 1 | 0 | 1 |
| Triodiasp | 11 | 65 | 3 | 0 |
This provides a list of each species and counts of occurrences in Both, CameraOnly, FieldOnly and Absent categories. We now calculate the camera detection rate for each species based on these data.
# first ensure/convert the data to numeric
species_counts$Absent <- as.numeric(as.character(species_counts$Absent))
species_counts$Both <- as.numeric(as.character(species_counts$Both))
species_counts$FieldOnly <- as.numeric(as.character(species_counts$FieldOnly))
species_counts$CameraOnly <- as.numeric(as.character(species_counts$CameraOnly))
# next calculate the total and number of valid, positive records of each species, which we assume here are counts that occur in either field surveys or both field surveys and camera surveys (the 'Both' category)
species_counts$Positives <- species_counts$Both + species_counts$FieldOnly
species_counts$Total <- species_counts$Absent + species_counts$Both + species_counts$FieldOnly + species_counts$CameraOnly
# rank the species by total positive count
species_counts<-species_counts[order(species_counts$Positives,decreasing = TRUE),]
# calcuate the frequency of each species as a percentage occurrence across plots
species_counts$Frequency <- 100*(species_counts$Positives / species_counts$Total)
# calculate the percentage of positive occurrences that were captured in the camera survey for each species #
species_counts$CameraDetectionRate <- 100-100*(species_counts$FieldOnly / species_counts$Positives)
# reorder the columns for ease of viewing and present the data
species_counts <- species_counts %>%
select(Species, Total, Positives,
Frequency, CameraDetectionRate,
Absent, Both,
CameraOnly, FieldOnly)
species_counts_gt_table <- gt(species_counts) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Counts all species")
)
species_counts_gt_table| Table 7 : Counts all species | ||||||||
|---|---|---|---|---|---|---|---|---|
| Species | Total | Positives | Frequency | CameraDetectionRate | Absent | Both | CameraOnly | FieldOnly |
| Triodiasp | 79 | 65 | 82.278481 | 100.00000 | 11 | 65 | 3 | 0 |
| AcacAneu | 79 | 26 | 32.911392 | 100.00000 | 50 | 26 | 3 | 0 |
| AlutMais | 79 | 23 | 29.113924 | 86.95652 | 52 | 20 | 4 | 3 |
| AlloDeca | 79 | 20 | 25.316456 | 100.00000 | 59 | 20 | 0 | 0 |
| GrevErio | 79 | 19 | 24.050633 | 94.73684 | 59 | 18 | 1 | 1 |
| CodoCoti | 79 | 13 | 16.455696 | 92.30769 | 66 | 12 | 0 | 1 |
| AcacLigu | 79 | 11 | 13.924051 | 63.63636 | 68 | 7 | 0 | 4 |
| EucaGamo | 79 | 8 | 10.126582 | 100.00000 | 71 | 8 | 0 | 0 |
| AcacMiny | 79 | 6 | 7.594937 | 83.33333 | 73 | 5 | 0 | 1 |
| EremLatr | 79 | 4 | 5.063291 | 50.00000 | 75 | 2 | 0 | 2 |
| AcacMait | 79 | 3 | 3.797468 | 66.66667 | 75 | 2 | 1 | 1 |
| DodoVisc | 79 | 3 | 3.797468 | 66.66667 | 76 | 2 | 0 | 1 |
| EucaOxy | 79 | 3 | 3.797468 | 66.66667 | 76 | 2 | 0 | 1 |
| GrevSten | 79 | 3 | 3.797468 | 66.66667 | 76 | 2 | 0 | 1 |
| NewcSpod | 79 | 3 | 3.797468 | 100.00000 | 75 | 3 | 1 | 0 |
| AcacAmmo | 79 | 2 | 2.531646 | 100.00000 | 77 | 2 | 0 | 0 |
| AcacMell | 79 | 2 | 2.531646 | 0.00000 | 77 | 0 | 0 | 2 |
| EremGlab | 79 | 2 | 2.531646 | 0.00000 | 77 | 0 | 0 | 2 |
| EremLong | 79 | 2 | 2.531646 | 100.00000 | 77 | 2 | 0 | 0 |
| GrevJunc | 79 | 2 | 2.531646 | 0.00000 | 77 | 0 | 0 | 2 |
| GyroTepp | 79 | 2 | 2.531646 | 100.00000 | 77 | 2 | 0 | 0 |
| SennArte | 79 | 2 | 2.531646 | 50.00000 | 77 | 1 | 0 | 1 |
| AcacPrui | 79 | 1 | 1.265823 | 100.00000 | 78 | 1 | 0 | 0 |
| DicrGile | 79 | 1 | 1.265823 | 100.00000 | 78 | 1 | 0 | 0 |
| ExocSpar | 79 | 1 | 1.265823 | 100.00000 | 78 | 1 | 0 | 0 |
| GyroRamo | 79 | 1 | 1.265823 | 100.00000 | 78 | 1 | 0 | 0 |
| RhagErem | 79 | 1 | 1.265823 | 0.00000 | 78 | 0 | 0 | 1 |
| SanLanc | 79 | 1 | 1.265823 | 0.00000 | 78 | 0 | 0 | 1 |
Now find species with more than 10 positive records and present them
# filter species with more than 10 positive records and present them
species_counts_gr_10 <- species_counts %>%
filter(Positives > 10) %>%
arrange(desc(Positives))
counts_gr_10_gt_table <- gt(species_counts_gr_10) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Counts of species in > 10 plots")
)
counts_gr_10_gt_table
# get the minimum camera detection rate for the first 6 species detected in more than 10 plots
min10 <- species_counts_gr_10 %>%
slice(1:6) %>% # Select the first 6 rows
summarise(MinCameraDetectionRate = min(CameraDetectionRate, na.rm = TRUE))
#cat(paste0("Minimum Camera Detection Rate: ", round(min10, 2)))
#cat("\n")
# get the minimum camera detection rate for Acacia ligulata, a species most frequently undetected; for the text below
AcacLigu_index <- which(species_counts_gr_10$Species == "AcacLigu")
Acli <- min(species_counts_gr_10$CameraDetectionRate[AcacLigu_index])
#cat(paste0("Acacia ligulata: ", round(Acli, 2)))
#cat("\n")| Table 8 : Counts of species in > 10 plots | ||||||||
|---|---|---|---|---|---|---|---|---|
| Species | Total | Positives | Frequency | CameraDetectionRate | Absent | Both | CameraOnly | FieldOnly |
| Triodiasp | 79 | 65 | 82.27848 | 100.00000 | 11 | 65 | 3 | 0 |
| AcacAneu | 79 | 26 | 32.91139 | 100.00000 | 50 | 26 | 3 | 0 |
| AlutMais | 79 | 23 | 29.11392 | 86.95652 | 52 | 20 | 4 | 3 |
| AlloDeca | 79 | 20 | 25.31646 | 100.00000 | 59 | 20 | 0 | 0 |
| GrevErio | 79 | 19 | 24.05063 | 94.73684 | 59 | 18 | 1 | 1 |
| CodoCoti | 79 | 13 | 16.45570 | 92.30769 | 66 | 12 | 0 | 1 |
| AcacLigu | 79 | 11 | 13.92405 | 63.63636 | 68 | 7 | 0 | 4 |
The objective of this analysis was to determine which species are most easily detected using camera images taken 3.2 m above the ground and in the centre of a 20 m diameter plot. The results show that species differed considerably in camera detectability, with the camera detection rate ranging from 0% to 100% depending on the species. Among common, widespread species or species groups (> 10 positive records or occurrences), Allocasuarina decaisneana, Acacia aneura, the Triodia species group (T. basedowii, T. schinzii, T. pungens), Aluta maisonneuvei, Codonocarpus continifolius, and Grevillea eriostachya were all detected in camera images 87% of the time or more. Only Acacia ligulata, many of which were dead across the study area, was an exception (64%).
Among less common species, morphologically distinct species such as Eucalyptus gamophylla and Newcastelia spodiotricha were always successfuly detected, and many other species were a majority of the time.
Conclusions : plant species detection using 360-degree cameras
Overall, plot-level species occurrence data were reliably extracted from camera images in this study. There was a strong relationship between species counts based on field survey and camera survey, albeit with a weak tendency for fewer species to be recorded in the camera survey. On average there was 84% agreement between field survey and camera survey species lists, although we note that the overall species diversity in the study area was relatively low, and apart from the large spinifex grasses (Triodia spp.) we did not focus on herbaceous species.
The most common species identified in the field survey but missed in the camera survey was Triodia schinzii, and it was not generally possible to distinguish between Triodia schinzii and Triodia pungens, even less than 5 m from the centre of the plot (i.e., where the photo was taken). However, this is true even under field conditions when plants are lacking floral structures. More species, particularly small taxa, seedlings and plants that retained no leaf material or species that co-occured with morphologically similar taxa were sometimes missed 5 m to 10 m from the plot centre. Overall, however, the most common species or species groups all had a camera detection rate of 86.96%, as did some less common taxa. Occasionally species were recorded in the camera survey but not the field survey, and since all species were perennial, these cases may be classified as field errors.
Collectively, we draw four conclusions from these data:
Camera data can be accurately used to generate plot-level species lists at the study site, although fewer species are missed within 5 m of the plot centre. Common and widespread species were all successfully detectable.
Morphologically distinct species were most easily detectable, and at greater distances from the camera. Morphologically similar species (e.g., spinifex grasses, Eremophila spp.) may need to be lumped together for subsequent analyses.
Reference photographs, ideally taken at different distances from the camera, should be provided for all target species to assist with species identification using the pannotator package.
The resolution of the current GoPro Max model may require collection of images closer to certain target taxa. This may involve, for example, collecting images closer to the ground (understory species), in the canopy (morphologically similar canopy taxa), and at multiple levels in both (structurally complex vegetation).
2. Spinifex grass (Triodia spp.) cover and dieback data
Total plot-level cover and mortality
The objective of the following analyses is to investigate the relationship between cover and mortality (dieback) estimates of spinifex (Triodia spp.) grasses across the validation plots in which cover and dieback estimates were generated both by field surveys and by using the pannotator package to score 360-degree images. We are particularly interested in specific size classes (if any) where discrepancies arise, and alterations to scoring methods that might improve accuracy if camera detection methods.
We first import the dataset spinifex_validation.csv and create the spinifex_data dataframe. The dataframe contains the following variables:
Column \[1\] (ID) = record identifier; \[2\] (Plot_NEW) = plot number; \[3\] (Location) = location of plot along transect (NE, NW, SE, SW corner or extra plot); \[4\] (Latitude) = latitude in decimal degrees; \[5\] (Longitude) = longitude in decimal degrees; \[6\] (Closest_image) = identity of closest 360-degree image (for pannotator annotation);
\[7\] (Field_Spinifex_size) = size class of spinifex identified in field survey (classes = 0-30 cm, >30-60, >60-100, >100-200, >200-300, >300-400, >400-500, >500 cm); \[8\] (Field_Spinifex_Cover) = field estimated cover (%); \[9\] (Field_Spinifex_%dead) = field percentage of spinifex cover estimate as dead; \[10\] (Camera_Spinifex_size) = size class of spinifex identified in camera survey; \[11\] (Camera_Spinifex_Cover) = camera estimated cover (%); \[12\] (Camera_Spinifex_%dead) = camera percentage of spinifex cover estimate as dead;
\[13\] (Cover_Cam_Field_Diff) = difference in cover estimate based on camera cover (%) - field cover (%); \[14\] (PercDead_Cam_Field_Diff) = difference in percentage of dead cover based on camera_%dead - field_%dead; \[15\] (Size_classes_field) = number of distinct spinifex size classes observed in plot based on field data; \[16\] (Size_classes_camera) = number of distinct spinifex size classes observed in plot based on cameradata; \[17\] (Misclassifications) = number of size classes in a given plot present in both field and camera surveys but differing by one or more size classes; \[18\] (Missed_classes) = number of size classes in a given plot missed completely in either in field and camera surveys; \[19\] (ErrorType1) = Error Type 1: ‘Both’ - size class observed in both field and camera surveys, ‘FieldOnly’ - size class observed only in field survey, CameraOnly’ - size class observed only in camera survey; \[20\] (ErrorType2) = Error Type 2: ‘No Error’ - size class classified the same on field and camera suveys, ‘Misclassed’ - size class misclassified in either field or camera suveys, ‘FieldError’ - error made on field estimate of size class, ‘CameraError’ - error made on camera estimate of size class; \[21\] (Cat1) = misclassified by one size class category; \[22\] (Distance) = distance from camera for error : ‘Less5’ - less than 5 m, ‘Great5’ - more than 5 m.
# read in the data file. This is the data for species and cover calibration
spinifex_data <- read_csv("./Data_files/spinifex_validation.csv", show_col_types = FALSE)In this section we take a similar approach to the evaluation of species identification described above. First we focus on overall plot cover and mortality data. These data are rows with the value of ‘Total’ in the Field_Spinifex_size and Camera_Spinifex_size columns.
# subset the data on only those containing "Total" plot cover values
spinifex_plot_total <- spinifex_data[spinifex_data$Field_Spinifex_size == 'TOTAL',]
# confirm the number of plots of spinifex cover data and print the result
cat("The number of rows in the subsetted dataframe is: ", nrow(spinifex_plot_total), "\n")The number of rows in the subsetted dataframe is: 73 We will begin by comparing estimates of mean field- and camera-based spinifex cover (the mean across all verification plots), and then the relationship between field- and camera-based spinifex cover.
# first ensure the data are numeric
spinifex_plot_total$Field_Spinifex_Cover <- as.numeric(as.character(spinifex_plot_total$Field_Spinifex_Cover))
spinifex_plot_total$Camera_Spinifex_Cover <- as.numeric(as.character(spinifex_plot_total$Camera_Spinifex_Cover))
# calculate the mean plot-level spinifex cover using field and camera survey data and print the results
Field_mean_sp <- summary(spinifex_plot_total$Field_Spinifex_Cover)
cat("The summary for mean total spinifex cover across plots based on field survey data is:\n")
print(Field_mean_sp)
cat("\n")
Camera_mean_sp <- summary(spinifex_plot_total$Camera_Spinifex_Cover)
cat("The summary for mean total spinifex cover across plots based on camera image data is:\n")
print(Camera_mean_sp)
cat("\n")The summary for mean total spinifex cover across plots based on field survey data is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
0.00 3.50 13.00 13.15 17.00 46.00
The summary for mean total spinifex cover across plots based on camera image data is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
0.00 4.50 14.50 14.63 22.50 46.00 Field survey and camera survey methods yielded similar mean (13.15% versus 14.63% respectively), maximum and minimum estimates of total plot cover across the entire data set (n = 73 plots). This indicates that overall bias in the two measurement techniques might be small, but it does not mean that predictability at the plot scale is also high. To assess whether field and camera survey techniques yield similar cover estimates at the plot scale we must look further at this relationship. We begin by looking at the correlation between plot-level estimates of total spinifex cover using both techniques (spinifex_plot_total\(Field_Spinifex_Cover and spinifex_plot_total\)Camera_Spinifex_Cover).
# plot the correlation between field cover and camera cover estimates with linear model
plot(spinifex_plot_total$Field_Spinifex_Cover, spinifex_plot_total$Camera_Spinifex_Cover,
main = "Spinifex cover, field vs camera, linear model",
xlab = "Total spinifex cover (Field Survey)",
ylab = "Total spinifex cover (Camera Survey)",
pch = 16, # Use filled circles as data points
col = "black", # Set point color
ylim = c(0, 60), # Set y-axis limits
xlim = c(0, 60)) # Set x-axis limits
abline(lm(Camera_Spinifex_Cover ~ Field_Spinifex_Cover, data = spinifex_plot_total), col = "red") # add a linear model to the plot
# view the linear model statistics
linearmodel <- lm(Camera_Spinifex_Cover ~ Field_Spinifex_Cover,
data=spinifex_plot_total)
# test whether the slope estimate differs from 1, which might indicate a bias towards lower values of camera cover for a given value of field cover
# 1. Get the summary of the model
modelsummary <- summary(linearmodel)
cat("Linear model summary of Camera_Spinifex_Cover ~ Field_Spinifex_Cover \n")
print(modelsummary)
# 2. Get the standard error of the slope
slope_se <- coef(modelsummary)["Field_Spinifex_Cover", "Std. Error"]
cat(paste0("Slope of the standard error: ", slope_se, "\n"))
# 3. Get the slope estimate
slope_estimate <- coef(modelsummary)["Field_Spinifex_Cover", "Estimate"]
cat(paste0("Slope estimater: ", slope_estimate, "\n"))
# 4. Calculate the Z test statistic (how many standard errors away from 1 is the slope estimate)
test_statistic <- (slope_estimate - 1) / slope_se
cat(paste0("Z score of the slope estimate: ", test_statistic, "\n"))
# 5. Calculate the p-value for the two-sided test
p_value <- 2 * (1 - pnorm(abs(test_statistic)))
# 6. Print the p-value
cat(paste0("The P value for testing deviation in the slope from 1 is: ", p_value, "\n"))Linear model summary of Camera_Spinifex_Cover ~ Field_Spinifex_Cover
Call:
lm(formula = Camera_Spinifex_Cover ~ Field_Spinifex_Cover, data = spinifex_plot_total)
Residuals:
Min 1Q Median 3Q Max
-15.915 -3.307 -1.695 3.013 22.555
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 3.30680 1.22624 2.697 0.00874 **
Field_Spinifex_Cover 0.86105 0.07105 12.119 < 2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 6.786 on 71 degrees of freedom
Multiple R-squared: 0.6741, Adjusted R-squared: 0.6695
F-statistic: 146.9 on 1 and 71 DF, p-value: < 2.2e-16
Slope of the standard error: 0.0710465545046707
Slope estimater: 0.861045180580407
Z score of the slope estimate: -1.95582770182694
The P value for testing deviation in the slope from 1 is: 0.0504854544045341Both indicate that field survey cover is strongly related (p < 5.8946282^{-19}) to camera survey cover. The linear model explains 67% of variability in the data, and the test for whether the slope of the relationship (0.86) differs from 1 is marginally significant (p = 0.05), which in this model indicates that camera cover scores may be slightly biased towards lower values than field survey estimates. Visually, however, the relationship appears to be non-linear, with stronger bias at higher values of field survey cover. Next we do a very simple investigation of this possible bias using a second-order polynomial function.
# generate a simple second-order polynomial to capture the possible decrease in camera cover estimate at higher values of field cover estimate
poly2model <- lm(Camera_Spinifex_Cover ~ Field_Spinifex_Cover + I(Field_Spinifex_Cover^2), data = spinifex_plot_total)
# generate a sequence of values spanning the range of Field_Spinifex_Cover
newdata <- data.frame(Field_Spinifex_Cover = seq(min(spinifex_plot_total$Field_Spinifex_Cover), max(spinifex_plot_total$Field_Spinifex_Cover), length.out = 100))
# use the model to predict Camera_Spinifex_Cover for these values
newdata$Camera_Spinifex_Cover <- predict(poly2model, newdata)
# plot the predictions
plot(Camera_Spinifex_Cover ~ Field_Spinifex_Cover, data = spinifex_plot_total,
main = "Spinifex cover, field vs camera, polynomial model",
xlab = "Total spinifex cover (Field Survey)",
ylab = "Total spinifex cover (Camera Survey)",
pch = 16, # Use filled circles as data points
col = "black", # Set point color
ylim = c(0, 60), # Set y-axis limits
xlim = c(0, 60)) # Set x-axis limits)
lines(Camera_Spinifex_Cover ~ Field_Spinifex_Cover, data = newdata, col = "red")
# view the second order polynomial model statistics
a2 <- summary(poly2model)
cat("Summary of polynomial model: \n")
print(a2)
cat("\n")
# calculate AIC for both models
aicl <- aic_linearmodel <- AIC(linearmodel)
aicp <- aic_poly2model <- AIC(poly2model)
# print the AIC values
cat(paste0("AIC for linearmodel: ", aic_linearmodel, "\n"))
cat(paste0("AIC for poly2model: ", aic_poly2model, "\n"))
# get the model adjusted R squared
R2 <- a2$adj.r.squared
cat("The model adjusted R^2 is: ", R2, "\n")Summary of polynomial model:
Call:
lm(formula = Camera_Spinifex_Cover ~ Field_Spinifex_Cover + I(Field_Spinifex_Cover^2),
data = spinifex_plot_total)
Residuals:
Min 1Q Median 3Q Max
-14.8427 -3.3427 -0.5873 2.2686 20.3047
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.587326 1.414317 0.415 0.67921
Field_Spinifex_Cover 1.410770 0.179514 7.859 3.32e-11 ***
I(Field_Spinifex_Cover^2) -0.015139 0.004591 -3.297 0.00154 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 6.358 on 70 degrees of freedom
Multiple R-squared: 0.7179, Adjusted R-squared: 0.7099
F-statistic: 89.09 on 2 and 70 DF, p-value: < 2.2e-16
AIC for linearmodel: 490.696986792716
AIC for poly2model: 482.157066085602
The model adjusted R^2 is: 0.7098859 The second order parameter explains a significant proportion of the data (p = 0.0015353) and the model adjusted R2 is slightly higher than the linear model (0.71). The lower AIC criterion for the polynomial model (490.7 vs. 482.16) also suggests that it provides a better fit to the relationship between field and camera cover estimates. Evaluation of the graphical relationship suggests that any bias towards lower camera cover scores occurs at field survey cover estimates of >30%, which is only a small minority of plots.
We now consider more closely the magnitude of differences in plot-level estimates of spinifex cover based on field survey and camera survey methods. The variable of interest is spinifex_plot_total\(Cover_Cam_Field_Diff which is the difference in total cover between the camera cover estimate (TCcam) and the field cover estimate (TCfield); ie. Cover_Cam_Field_Diff = TCcam - TCfield (the two variables being spinifex_plot_total\)Camera_Spinifex_Cover and spinifex_plot_total$Field_Spinifex_Cover).
# determine the mean difference in plot level field and camera cover estimates
a <- summary(spinifex_plot_total$Cover_Cam_Field_Diff)
cat("The summary for mean cover difference based on all data is:\n")
print(a)
cat("\n")
# create the histogram:
hist_data_all <- hist(spinifex_plot_total$Cover_Cam_Field_Diff,
breaks = seq(-50, 50, by = 5),
main = "Histogram of cover estimate differences",
xlab = "Cover difference %",
ylab = "Frequency")
The summary for mean cover difference based on all data is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
-19.000 -1.000 0.000 1.479 4.000 23.500 # determine the difference for only plots that contain Triodia (i.e., excluding plots with zero spinifex cover in both field and camera surveys)
spinifex_plot_total_gr0 <- subset(spinifex_plot_total, Field_Spinifex_Cover > 0 | Camera_Spinifex_Cover > 0)
b <- summary(spinifex_plot_total_gr0$Cover_Cam_Field_Diff)
cat("The summary for mean cover difference based on only plots containing spinifex is:\n")
print(b)
cat("\n")
# create the histogram
hist_data_gr0 <- hist(spinifex_plot_total_gr0$Cover_Cam_Field_Diff,
breaks = seq(-50, 50, by = 5),
main = "Histogram of cover estimate differences \n(plots containing spinifex)",
xlab = "Cover difference %",
ylab = "Frequency")
The summary for mean cover difference based on only plots containing spinifex is:
Min. 1st Qu. Median Mean 3rd Qu. Max.
-19.000 -1.500 0.000 1.742 5.375 23.500 The data confirm that differences in total spinifex cover estimates based on field survey and camera survey data are tightly clustered near a mean of zero and an approximately normal distribution. Differences can be as large as 20% but the majority are in the range of -10% to +10%.
We must now also consider differences in the percentage of spinifex cover that is dead (suffering recent mortality/dieback). Here we provide an abbreviated analysis based on key results.
Field_mean <- summary(spinifex_plot_total$`Field_Spinifex_%dead`)
cat("The summary for mortality (% dead cover) across all plots based on field survey data is:\n")
print(Field_mean)
cat("\n")
Camera_mean <- summary(spinifex_plot_total$`Camera_Spinifex_%dead`)
cat("The summary for mortality (% dead cover) across all plots based on camera image data is:\n")
print(Camera_mean)
cat("\n")The summary for mortality (% dead cover) across all plots based on field survey data is:
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
0.00 0.10 10.14 25.64 47.44 100.00 11
The summary for mortality (% dead cover) across all plots based on camera image data is:
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
0.000 4.364 24.773 29.968 45.903 100.000 11 The means and spread of data are similar.
# plot the relationship between mortality (%dead) estimates from both field and camera surveys
plot(spinifex_plot_total$`Field_Spinifex_%dead`, spinifex_plot_total$`Camera_Spinifex_%dead`,
main = "Percent dead cover, field vs camera, linear model",
xlab = "Percent dead cover (Field Survey)",
ylab = "Percent dead cover (Camera Survey)",
pch = 16, # Use filled circles as data points
col = "black", # Set point color
ylim = c(0, 100), # Set y-axis limits
xlim = c(0, 100)) # Set x-axis limits
abline(lm(`Camera_Spinifex_%dead` ~ `Field_Spinifex_%dead`, data = spinifex_plot_total), col = "red") # add a linear model to the plot
# view the linear model statistics
linearmodel <- lm(`Camera_Spinifex_%dead` ~ `Field_Spinifex_%dead`, data = spinifex_plot_total)
a3 <- summary(linearmodel)
cat("Summary of the linear model: Camera_Spinifex_%dead ~ Field_Spinifex_%dead \n")
print(a3)
cat("\n")
# print the P value of the model
pvalue <- a3$coefficients[2,4]
cat("The P value of the model is: ", pvalue, "\n")Summary of the linear model: Camera_Spinifex_%dead ~ Field_Spinifex_%dead
Call:
lm(formula = `Camera_Spinifex_%dead` ~ `Field_Spinifex_%dead`,
data = spinifex_plot_total)
Residuals:
Min 1Q Median 3Q Max
-34.622 -10.594 -6.637 8.417 89.406
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 10.59406 3.18337 3.328 0.0015 **
`Field_Spinifex_%dead` 0.75574 0.08101 9.329 2.77e-13 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 19 on 60 degrees of freedom
(11 observations deleted due to missingness)
Multiple R-squared: 0.5919, Adjusted R-squared: 0.5851
F-statistic: 87.03 on 1 and 60 DF, p-value: 2.772e-13
The P value of the model is: 2.771671e-13 The linear model is highly significant (p = 2.7716709^{-13})), and explains 59% of variation in the data. We now look at the spread in the data:
# create the histogram
hist_data <- hist(spinifex_plot_total$PercDead_Cam_Field_Diff,
breaks = seq(-100, 100, by = 10),
main = "Histogram of differences in mortality estimates",
xlab = "Difference in estimated mortality (%)",
ylab = "Frequency")
The data are again centrally distributed with a mean near zero, but the spread in the data are larger than for live cover. The majority of errors are between -20% to +20%, but some estimates vary by up to ca. 50-60%.
Within-plot differences in cover and mortality by size class
The object of the following analyses is to investigate differences in classification among size classes within plots. First we consider differences in the mean number of size classes per plot based on the field and camera data, along with the number of misclassified size classes (one size class category differences) and missed classes (size classes present in either field or camera data but not both).
SCField_mean <- summary(spinifex_plot_total$Size_classes_field)
cat("Summary - number of spinifex size classes per plot based on field survey data:\n")
print(SCField_mean)
cat("\n")
SCCamera_mean <- summary(spinifex_plot_total$Size_classes_camera)
cat("Summary - number of spinifex size classes per plot based on camera survey data:\n")
print(SCCamera_mean)
cat("\n")
SCMisclass_mean <- summary(spinifex_plot_total$Misclassifications)
cat("Summary - spinifex size class misclassifications per plot:\n")
print(SCMisclass_mean)
cat("\n")
SCMissed_class_mean <- summary(spinifex_plot_total$Missed_classes)
cat("Summary - number of spinifex size classes missed per plot in either survey:\n")
print(SCMissed_class_mean)
cat("\n")Summary - number of spinifex size classes per plot based on field survey data:
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
0.000 1.000 2.000 1.958 3.000 5.000 1
Summary - number of spinifex size classes per plot based on camera survey data:
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
0.000 1.000 2.000 2.083 3.000 5.000 1
Summary - spinifex size class misclassifications per plot:
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
0.0000 0.0000 0.0000 0.3226 1.0000 2.0000 11
Summary - number of spinifex size classes missed per plot in either survey:
Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
0.0000 0.0000 0.0000 0.3226 1.0000 2.0000 11 Overall, the mean number of spinifex size classes per plot recorded in field (1.96) and camera (2.08) data were very similar, and the mean number of misclassifications (0.32) and missed size class categories (0.32) were both very low.
We now generate a new dataframe that contains cover and mortality data for all size classes in each plot, removing plot totals. We then determine the number of occasions in which a given spinifex size class was observed in both field and camera data (Both), or in only field (FieldOnly) or camera (CameraOnly) data.
# Count occurrences of each ErrorType1 for non-TOTAL Field_Spinifex_size
size_class_data <- spinifex_data %>%
filter(Field_Spinifex_size != "TOTAL",
ErrorType1 %in% c("Both", "CameraOnly", "FieldOnly")) %>%
count(ErrorType1, sort = TRUE)
#View(size_class_data)
size_class_data_gt_table <- gt(size_class_data) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Spinifex size class data")
)
size_class_data_gt_table
# Calculate total counts and percentages for each error type
counttotal <- sum(size_class_data$n)
pcnt_both <- round(100 * (size_class_data$n[size_class_data$ErrorType1 == "Both"] / counttotal),2)
pcnt_camera <- round(100 * (size_class_data$n[size_class_data$ErrorType1 == "CameraOnly"] / counttotal),2)
pcnt_field <- round(100 * (size_class_data$n[size_class_data$ErrorType1 == "FieldOnly"] / counttotal),2)
#print(paste0("Percent in both: ", pcnt_both))
#print(paste0("Percent in camera: ", pcnt_camera))
#print(paste0("Percent in field: ", pcnt_field))| Table 9 : Spinifex size class data | |
|---|---|
| ErrorType1 | n |
| Both | 123 |
| CameraOnly | 30 |
| FieldOnly | 22 |
The data show that 70.3% of spinifex size classes were recorded in both Camera and Field surveys, while 29.7% occurred only in one survey. Of these, more were recorded only in the camera survey (n = 30) than in the field survey (n = 22).
We now look at how these differences are distributed through spinifex size classes. In this study we used the following size classes: 0-30 cm, >30-60, >60-100, >100-200, >200-300,>300-400, >400-500, >500cm.
# first reorder the size classes from smallest to largest
levels_order <- c("0-30", ">30-60", ">60-100", ">100-200", ">200-300", ">300-400" , ">400-500", ">500")
# convert Field_Spinifex_size to a factor and specify the order of the levels
spinifex_data$Camera_Spinifex_size <- factor(spinifex_data$Camera_Spinifex_size, levels = levels_order)
# display the data
cross_matrix <- as.data.frame.matrix(table(spinifex_data$Camera_Spinifex_size, spinifex_data$ErrorType1)) %>%
rownames_to_column(var = "Size class") %>%
arrange(factor(`Size class`, levels = levels_order))
gt_table <- gt(cross_matrix) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Spinifex size class data & error type 1")
)
gt_table| Table 10 : Spinifex size class data & error type 1 | |||
|---|---|---|---|
| Size class | Both | CameraOnly | FieldOnly |
| 0-30 | 32 | 9 | 2 |
| >30-60 | 34 | 4 | 5 |
| >60-100 | 30 | 8 | 1 |
| >100-200 | 18 | 3 | 8 |
| >200-300 | 9 | 3 | 5 |
| >300-400 | 0 | 1 | 0 |
| >400-500 | 0 | 2 | 0 |
| >500 | 0 | 0 | 1 |
There appears to be a tendency for smaller size classes to be recorded solely on camera images and the largest size classes to be recorded only in the field surveys.
We then attempted to determine, by comparing field notes and camera data, where these discrepancies occurred. Here, errors were classified according to: Camera error: the size class was present but definitely missed in the camera survey, Field error: the size class was present but definitely missed in the field survey, or Misclassified : the size class was seen in both but allocated to different size classes.
# Define the order of levels
levels_order <- c("0-30", ">30-60", ">60-100", ">100-200", ">200-300", ">300-400", ">400-500", ">500")
# Ensure the Field_Spinifex_size column is a factor with the specified order
spinifex_data$Field_Spinifex_size <- factor(spinifex_data$Field_Spinifex_size, levels = levels_order)
# Create the cross matrix and order it according to levels_order
cross_matrix2 <- as.data.frame.matrix(table(spinifex_data$Field_Spinifex_size, spinifex_data$ErrorType2)) %>%
rownames_to_column(var = "Size class") %>%
arrange(factor(`Size class`, levels = levels_order))
# generate the cross matrix table
cross_matrix2 <- as.data.frame.matrix(table(spinifex_data$Field_Spinifex_size, spinifex_data$ErrorType2)) %>%
rownames_to_column(var = "Size class")
size_class_gt_table <- gt(cross_matrix2) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Spinifex size class data & error type 2")
)
size_class_gt_table
# Count the number of different error types and do not sort them
Error2_summary <- spinifex_data %>%
count(ErrorType2, sort = TRUE)
# #Extract counts by category name
Camerror <- Error2_summary$n[which(Error2_summary$ErrorType2 == "CameraError")]
Fielderror <- Error2_summary$n[which(Error2_summary$ErrorType2 == "FieldError")]
Miserror <- Error2_summary$n[which(Error2_summary$ErrorType2 == "Misclassed")]
No_error <- Error2_summary$n[which(Error2_summary$ErrorType2 == "No Error")]
Total <- sum(Error2_summary$n)
# Optionally, you can print the counts for verification
# cat("Camera Errors: ", Camerror, "\n")
# cat("Field Errors: ", Fielderror, "\n")
# cat("Misclassified: ", Miserror, "\n")
# cat("No Errors: ", No_error, "\n")
# cat("Total: ", Total, "\n")| Table 11 : Spinifex size class data & error type 2 | ||||
|---|---|---|---|---|
| Size class | CameraError | FieldError | Misclassed | No Error |
| 0-30 | 2 | 6 | 3 | 32 |
| >30-60 | 4 | 3 | 2 | 34 |
| >60-100 | 0 | 3 | 6 | 30 |
| >100-200 | 3 | 1 | 7 | 18 |
| >200-300 | 2 | 1 | 5 | 9 |
| >300-400 | 0 | 0 | 1 | 0 |
| >400-500 | 0 | 1 | 1 | 0 |
| >500 | 0 | 0 | 1 | 0 |
The validation data show that specific size classes were missed in either the camera survey (4 % of cases) or the field survey (6 % of cases), which indicates that camera surveys were actually more reliable in some circumstances than field surveys. In another 26 cases (10 %) size classes were simply misclassified by one size class category.
Conclusions : spinifex size class and mortality detection using 360-degree cameras
Overall, analysis of the validation data set suggests that spinifex cover and mortality data extracted from 360-degree images performs well when compared with data collected during field surveys. This was true even given the large size of the plots (20 m diameter) and the fact that only one image was taken in the centre of each plot. Over a large sample of plots the mean plot-level cover and mortality rate (percentage of dead material) using both techniques were very similar and had errors that were approximately normally distributed with a mean of near zero, indicating minimal bias using camera images for data collection. Generally, plot-level cover estimates varied by ~ +10% to -10%, and percentage mortality estimates by ~ +20% to -20%, although differences in a small minority of plots were larger. On average, the mean number of spinifex size classes per plot was extremely similar based in field data and camera data, and far in excess of the number of misclassified or missed size classes.
At the plot level, we used a fine-grained cover size scale (8 classes), particularly for tussocks <1m across (0-30 cm, >30-60 cm, and >60-100cm). Despite this, 70% of size classes were correctly identified in both camera and field survey data. A further 9% were only identified in camera data, due to field errors in data collection, thus ~80% of size classes were correctly identified using camera surveys. The remaining discrepancies were due to size class estimates developed using camera and field survey techniques differing by a single size class. There was a weak tendency for smaller size classes to be recorded solely in camera images and the largest size classes to be recorded only in the field surveys.
Collectively, we draw four conclusions from these data:
1) Mean cover and percent mortality estimates for spinifex grasses across plots based on camera images were very similar to those based on field survey data, and there was no directional biases in cover and dieback differences. Therefore, for large samples, mean estimates for both parameters using camera images are likely to be accurate.
2) At the within-plot level, camera and field survey approaches yielded similar numbers of size classes per plot, and, overall, the number of size classes misclassified or missed completely was small. Overall, about 80% of specific size classes identified using camera surveys were either verified by field data or were correct when field data were incorrect. About 6% of size classes were definitely due to errors in data collected from 360-degree images. Thus, camera surveys generate data at least, or possibly slightly better than, field surveys based on visual cover and dieback estimates.
3) These data and the obvious relationship between the size classes of spinifex present in a given area and fire history (see accompanying paper) indicate that 360-degree images are a reliable source of cover and dieback data. However, the discrepancies among size class data estimates indicate that there is still room for improvement. Therefore, for the transect study reported in the accompanying paper we decide to undertake the following changes:
a) reduce the size of the study plot from 20 m to 10 m diameter, where camera images are most useful,
b) overlay plot boundaries and cover estimate markers onto 360-degree images, and
c) develop help files that contain images in which cover and dieback estimates were determined with a very high degree of accuracy using the ImageJ software package.
Ultimately, at the 10 m diameter scale spinifex cover and dieback estimates are likely to be as good or better than those developed using standard visual field survey estimates. We recommend that all studies collect similar verification data, and carefully consider optimal camera positioning for different taxa and ecological situations. Calibration or validation studies may benefit from even more intensive field survey techniques, for example line intercept or point-based sampling. Ultimately, however, it is our belief that cover or dieback estimates generated using imaging software (such as ImageJ), while time-consuming to collect without automation, will ultimately be far more accurate than those generated using even more intensive field survey estimates.
3. Desert oak crown health
The objective of the following analyses is to investigate variation in crown health and size class data for desert oak (Allocasuarina decaisneana) extracted from camera images by three different scorers using the pannotator package. Inter-observer variability is a problem in many scientific disciplines and our intention was to determine the scoring protocols and help files that minimise these differences.
Here we compare three scorers: 1) one present at the time when images were collected in cameras and is familiar with the study site and associated vegetation (scorer 1), 2) two with no experience at the study site or in scoring crown health (scorers 2-3). Scorer 1 had access to a rudimentary help file (fewer images), a less refined size class scoring system (no seedling size class), and a single worked example for calculating the crown health score. All camera images were collected from the 4 km-long transect that forms the basis of the study reported in this accompanying paper.
We first import the required datasets: scorer_oak_1.csv, scorer_oak_2.csv, and scorer_oak_3.csv.
Each dataframe contains the following variables:
Column \[1\] (user) = scorer number; \[2\] (timestamp) = image timestamp, \[3\] (leaflet_id) = leaflet ID; \[4\] (imagefile) = related image file name; \[5\] (feature_type) = feature type (marker or image marker); \[6\] (radius) = radius geometry (not applicable); \[7\] (geometry) = geometry in Well-Known Text (WKT) format; \[8\] (species) = species name & size/impact class as per dropdown menu; \[9\] (Size_class) = Size class (seedling, juvenile, early pole, late pole, early mature, mature); \[10\] (veg_layer) = vegetation level, here Allocasuarina decaisneana = tree; \[11-13\] (dd1-dd3) = crown health estimate components; \[14\] (plot) = image plot; \[15\] (plot_name_new) = plot names assigned in study ; \[16\] (frame) = frame of video image file was taken from.
Upload oak annotation files and add plot and frame numbers
# read in the annotation files
scorer_1 <- read.csv("./Data_files/scorer_oak_1.csv")
scorer_2 <- read.csv("./Data_files/scorer_oak_2.csv")
scorer_3 <- read.csv("./Data_files/scorer_oak_3.csv")Load scorer 1 data
# set the geometry
scorer_1 <- st_as_sf(scorer_1, wkt = "geometry", crs = 4326)
#mapview(scorer_1, width = "100%", na.rm = TRUE)Load scorer 2 data
# set the geometry
scorer_2 <- st_as_sf(scorer_2, wkt = "geometry", crs = 4326)
#mapview(scorer_2, width = "100%", na.rm = TRUE)Load scorer 3 data
# set the geometry
scorer_3 <- st_as_sf(scorer_3, wkt = "geometry", crs = 4326)
#mapview(scorer_3, width = "100%", na.rm = TRUE)These maps confirm that data for each of the scorers was collected from the same transect at Uluru-Kata Tjuta NP.
Calculate the crown score and mean values per size class
Crown health is simply defined as Crown Health Score = dd1 x dd2 x dd3; values range from 100 (perfect health) to 0 (dead). Here dd1 = the assessable crown (AC, as a percentage of the potential crown that contains live or dead tertiary branches or smaller), dd2 = the crown extent (CE, which is the percentage of the assessable crown that contains leaves), and dd3 = the crown density (CD; the percentage of skylight blocked by leaves within the crown extent). AC, CE and CD were assigned percentages using the scale SHC= 0, >0-5, >5-20, >20-35, >35-50, >50-65, >65-80, >80-95, >95-<100, 100%, and scale midpoints used for calculations. Further details and examples are provided in the accompanying paper and Supporting Information 6).
# replace all -999 values (used to signify NA) with zero. The calculation works since when dd2 or dd3 contain -999 the previous dd variable is necessarily zero.
scorer_1$dd3 <- replace(scorer_1$dd3, scorer_1$dd3 == -999, 0)
scorer_2$dd3 <- replace(scorer_2$dd3, scorer_2$dd3 == -999, 0)
scorer_3$dd3 <- replace(scorer_3$dd3, scorer_3$dd3 == -999, 0)
scorer_1$dd2 <- replace(scorer_1$dd2, scorer_1$dd2 == -999, 0)
scorer_2$dd2 <- replace(scorer_2$dd2, scorer_2$dd2 == -999, 0)
scorer_3$dd2 <- replace(scorer_3$dd2, scorer_3$dd2 == -999, 0)
# develop a function to calculate Crown health scores in each dataframe and then reorder the size classes. Reordering the size classes is necessary for creating plots below
calculate_scores <- function(df) {
df$Crown_score <- df$dd1/100 * df$dd2/100 * df$dd3/100
df$Size_class_Reorder <- factor(df$Size_class, c("Seedling", "Juvenile", "Early_Pole", "Late_Pole", "Early_Mature", "Mature"))
return(df)
}
# use above function on scorer_1, scorer_2 and scorer_3
scorer_1 <- calculate_scores(scorer_1)
scorer_2 <- calculate_scores(scorer_2)
scorer_3 <- calculate_scores(scorer_3)Now we can compare mean crown health scores across different scorers.
# function to calculate mean crown scores in each size class
mean_and_count_table <- function(data, score_col, size_col, scorer_name) {
mean_table <- data %>%
group_by(!!sym(size_col)) %>%
summarize(
mean = mean(!!sym(score_col), na.rm = TRUE),
count = n()
) %>%
mutate(Scorer = scorer_name) %>%
rename(Category = !!sym(size_col))
return(mean_table)
}
table_scorer_1 <- mean_and_count_table(scorer_1, "Crown_score", "Size_class_Reorder", "Scorer 1")
#View(table_scorer_1)
table_scorer_2 <- mean_and_count_table(scorer_2, "Crown_score", "Size_class_Reorder", "Scorer 2")
#View(table_scorer_2)
table_scorer_3 <- mean_and_count_table(scorer_3, "Crown_score", "Size_class_Reorder", "Scorer 3")
#View(table_scorer_3a)
combined_table <- rbind(table_scorer_1, table_scorer_2, table_scorer_3)
#View(combined_table)
reshaped_table <- combined_table %>%
st_set_geometry(NULL) %>% # Remove the sf geometry
gather(key = "Statistic", value = "Value", -Category, -Scorer) %>%
unite("Scorer_Statistic", Scorer, Statistic, sep = "_") %>%
spread(key = "Scorer_Statistic", value = "Value")
# Print the reshaped table using gt
scorers_gt_table <- gt(reshaped_table) %>%
tab_options(
table.align = "center",
table.font.size = px(10),
column_labels.font.size = px(16),
heading.align = "center",
heading.title.font.size = px(20)
) %>%
opt_row_striping() %>%
tab_header(
title = table_caption("Crown scores from different scorers")
)
scorers_gt_table
# we also get the mean crown health score for each scorer
s1mean <- summary(scorer_1$Crown_score)
#print(s1mean)
s2mean <- summary(scorer_2$Crown_score)
#print(s2mean)
s3mean <- summary(scorer_3$Crown_score)
#print(s3mean)| Table 12 : Crown scores from different scorers | ||||||
|---|---|---|---|---|---|---|
| Category | Scorer 1_count | Scorer 1_mean | Scorer 2_count | Scorer 2_mean | Scorer 3_count | Scorer 3_mean |
| Seedling | NA | NA | 4 | 0.2437500 | 2 | 0.3625000 |
| Juvenile | 9 | 0.6034549 | 19 | 0.3264803 | 26 | 0.5147356 |
| Early_Pole | 7 | 0.5332500 | 16 | 0.6268896 | 13 | 0.6940385 |
| Late_Pole | 4 | 0.5954453 | 35 | 0.5449272 | 25 | 0.6838250 |
| Early_Mature | 25 | 0.5206769 | 52 | 0.4669056 | 38 | 0.6420189 |
| Mature | 51 | 0.4164403 | 20 | 0.5977703 | 31 | 0.6274970 |
These data indicate that crown health scores were similar for scorers 1 (0.477) and 2 (0.497) but and about 0.1 units higher for scorer 3 (0.623), mainly due to a lack of low-scoring trees. Scores for pole- and mature-sized trees were generally consistent across the three scorers, although again scores provided by scorer 3 tended to be higher than those provided by the other scorers. Scores differed in seedling and juvenile classes, although the number of individuals used to determine these scores varied significantly. Data provided by scorer 1 indicated that smaller seedlings (<1 m tall) needed to separated from juveniles (1- 4 m tall), and that a greater range of visual examples of all size classes was required in the help files. Incorporation of these changes greatly improved the agreement across size classes by scorers 2 and 3.
Now we want to determine whether the spatial structure of size class and crown health data differed significantly among scorers. Maps are visualised using mapview and images saved using the ‘mapshot’ function.
Scorer 1:
scorer_1 <- st_as_sf(scorer_1, wkt = "geometry",crs = 4326)
s1_p <- mapview(scorer_1, zcol = "Size_class_Reorder", cex = 8, alpha = 0.9, alpha.regions = 0.9, layer.name = "Size_class", col.regions = c("green4", "green4","yellow", "yellow", "red4", "red4"),na.rm = TRUE, basemaps = c("Esri.WorldImagery"))
#print(s1_p)
mapviewOptions(basemaps = c("Esri.WorldImagery"),
raster.palette = grey.colors,
vector.palette = colorRampPalette(c("red", "yellow", "green2")),
na.color = "gray",
layers.control.pos = "topright")
s1_q <- mapview(scorer_1, zcol = "Crown_score", layer.name = "Crown_score", cex = 8, alpha = 0.9, alpha.regions = 0.9, na.rm = TRUE, legend = TRUE, at = seq(0, 1, 0.1))
#print(s1_q)
# map crown scores of mature and early_mature trees
Mature <- subset(scorer_1, scorer_1$Size_class == "Mature" | scorer_1$Size_class == "Early_Mature")
s1_z <- mapview(Mature, zcol = "Crown_score", layer.name = "Crown_score", cex = 8, alpha = 0.9, alpha.regions = 0.9, na.rm = TRUE)
#print(s1_z)Scorer 2:
scorer_2 <- st_as_sf(scorer_2, wkt = "geometry",crs = 4326)
s2_p <- mapview(scorer_2, zcol = "Size_class_Reorder", cex = 8, alpha = 0.9, alpha.regions = 0.9, layer.name = "Size_class", col.regions = c("green4", "green4","yellow", "yellow", "red4", "red4"),na.rm = TRUE)
#print(s2_p)
mapviewOptions(basemaps = c("Esri.WorldImagery"),
raster.palette = grey.colors,
vector.palette = colorRampPalette(c("red", "yellow", "green2")),
na.color = "gray",
layers.control.pos = "topright")
s2_q <- mapview(scorer_2, zcol = "Crown_score", layer.name = "Crown_score", cex = 8, alpha = 0.9, alpha.regions = 0.9, na.rm = TRUE, legend = TRUE, at = seq(0, 1, 0.1))
#print(s2_q)
# map crown scores of mature and early_mature trees
Mature <- subset(scorer_2, scorer_2$Size_class == "Mature" | scorer_2$Size_class == "Early_Mature")
s2_z <- mapview(Mature, zcol = "Crown_score", layer.name = "Crown_score", cex = 8, alpha = 0.9, alpha.regions = 0.9, na.rm = TRUE)
#print(s2_z)Scorer 3:
s3_p <- mapview(scorer_3, zcol = "Size_class_Reorder", cex = 8, alpha = 0.9, alpha.regions = 0.9, layer.name = "Size_class", col.regions = c("green4", "green4","yellow", "yellow", "red4", "red4"),na.rm = TRUE)
#print(s3_p)
mapviewOptions(basemaps = c("Esri.WorldImagery"),
raster.palette = grey.colors,
vector.palette = colorRampPalette(c("red", "yellow", "green2")),
na.color = "gray",
layers.control.pos = "topright")
s3_q <- mapview(scorer_3, zcol = "Crown_score", layer.name = "Crown_score", cex = 8, alpha = 0.9, alpha.regions = 0.9, na.rm = TRUE, legend = TRUE, at = seq(0, 1, 0.1))
#print(s3_q)
# map crown scores of mature and early_mature trees
Mature <- subset(scorer_3, scorer_3$Size_class == "Mature" | scorer_3$Size_class == "Early_Mature")
s3_z <- mapview(Mature, zcol = "Crown_score", layer.name = "Crown_score", cex = 8, alpha = 0.9, alpha.regions = 0.9, na.rm = TRUE)
#print(s3_z)Compare Scorers using mapview
Now we can compare the scorers side by side. First size classes:
# Synchronize map views in html version
# sync(s1_p@map %>%
# addControl(html = "Scorer 1", position = "bottomright"),
# s2_p@map %>%
# addControl(html = "Scorer 2", position = "bottomright"),
# s3_p@map %>%
# addControl(html = "Scorer 3", position = "bottomright"))Now crown scores:
# sync(s1_q@map %>%
# addControl(html = "Scorer 1", position = "bottomright"),
# s2_q@map %>%
# addControl(html = "Scorer 2", position = "bottomright"),
# s3_q@map %>%
# addControl(html = "Scorer 3", position = "bottomright"))Now mature only:
# sync(s1_z@map %>%
# addControl(html = "Scorer 1 - Mature", position = "bottomright"),
# s2_z@map %>%
# addControl(html = "Scorer 2 - Mature", position = "bottomright"),
# s3_z@map %>%
# addControl(html = "Scorer 3 - Mature", position = "bottomright"))For convenience maps for size class and crown score have been compiled into single .png file:
Comparison of desert oak size class and crown health scores for 3 scorers. Examples of areas recently burnt (Burnt) and not recently burnt (Unburnt) are provided in the top left panel. In the top right panel F = areas apparently impacted severely by fire, D = areas not or minimally affected by fire, likely impact caused by drought.